becoming data-driven - machine learning @ xing marketing solutions
TRANSCRIPT
Becoming Data-drivenML @ XING Marketing Solutions
Big Data World Frankfurt
November 29th, 2017
Dr. Stefan Kühn
Dr. Stefan Kühn - Becoming Data-driven: Machine Learning @ XING Marketing Solutions 2
werben.xing.com
Dr. Stefan Kühn - Becoming Data-driven: Machine Learning @ XING Marketing Solutions 3
This talk is about
werben.xing.com
Dr. Stefan Kühn - Becoming Data-driven: Machine Learning @ XING Marketing Solutions 4
Create ads @ www.xing.com/xas/
Native Advertising• Different placements
• Multiple Ad Types
• Events• Groups• Jobs• User• BusinessPages• Websites• Video• …
Dr. Stefan Kühn - Becoming Data-driven: Machine Learning @ XING Marketing Solutions 5
AdManagerFrom Heuristics to Algorithms
Second Bid Auction
Dr. Stefan Kühn - Becoming Data-driven: Machine Learning @ XING Marketing Solutions 6
We focus on• Relevance
• Predict expected Clickrate = eCTR
• Revenue• Predict expected
Revenue = eRPI
• Features• Targeting• Ad-related data• User-related data• Time• Channel• …
Heuristics – Naïve Bayes (not quite)
Why?• Easy to implement
• No theoretic background needed
• No additional toolstack
• Can be implemented by Software Engineers
Dr. Stefan Kühn - Becoming Data-driven: Machine Learning @ XING Marketing Solutions 7
Why not?• Hard to optimize
• No theoretic guarantuees
• Toolstack limited w.r.t advanced methods
• Cannot be re-used by Data Scientist
Imprecise predictions lead to suboptimal business decisions
“Visible” costs are low “Invisible” costs are higher
Algorithms – Collaborative Filtering
Why not?• Significant implementation effort
• Complex theory
• New and unknown toolstack requires training and learning time
• Cannot be implemented by Software Engineers alone
Dr. Stefan Kühn - Becoming Data-driven: Machine Learning @ XING Marketing Solutions 8
Why?• Allows for ongoing optimization
• Theoretic guarantuees are a prerequisite for reasoning, proper evaluation and testing
• Modern tooling enables learning from much more data
• Additional Data Science and Engineering skills enhance the team capabilities in many ways
Short term savingsLong term benefits
Dr. Stefan Kühn - Becoming Data-driven: Machine Learning @ XING Marketing Solutions 9
Advanced Delivery Pipeline
Separation of Concerns
Dr. Stefan Kühn - Becoming Data-driven: Machine Learning @ XING Marketing Solutions 10
Collaborative Filtering for Recommendations
Predict approximate scores for empty spots based on similarities between users and items
Dr. Stefan Kühn - Becoming Data-driven: Machine Learning @ XING Marketing Solutions 11
Matrix Factorization
Dr. Stefan Kühn - Becoming Data-driven: Machine Learning @ XING Marketing Solutions 12
Short term savings versus long term benefits
0
20
40
60
80
100
120
140
Revenue per Impression [RPI]
Before ADP Start of ADP ADP today
+11%
+30%
• Development time 4 month (small team)
• In production for 4 months now and more to come
Dr. Stefan Kühn - Becoming Data-driven: Machine Learning @ XING Marketing Solutions 13
AdManagerMore to come
Algorithms – The Next Level
More Data = New Features• Natural Language Processing - Matching user
interests and ad descriptions
• Social Network Analysis - Recommendations based on interactions in the user’s network
• Interaction with other content
Dr. Stefan Kühn - Becoming Data-driven: Machine Learning @ XING Marketing Solutions 14
New Methods = Better Predictions• Multiple methods in parallel (Multi-Armed
Bandit)
• Multiple theoretical approaches (LogReg, Tree-based)
• Ensembles
New data dimensions require Big Data solution
New algorithmic dimensions require powerful distributed computing system
Algorithms – The Next Level
Dr. Stefan Kühn - Becoming Data-driven: Machine Learning @ XING Marketing Solutions 15
New data dimensions require Big Data solution
New algorithmic dimensions require powerful distributed computing system
We are already prepared for that!
Thank you for your attention.
www.xing.com
Dr. Stefan KühnSenior Data Scientist – XING Markting Solutions [email protected]/profile/Stefan_Kuehn46